Full Story w/ Open Brain Agent Memory: https://natesnewsletter.substack.com/p/openclaw-agent-runtime-model-swapping?r=1z4sm5&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true ___________________ What's really happening inside OpenClaw when everyone is arguing about the model layer but missing that the runtime itself changed shape in April? The common story is about Anthropic versus OpenAI and subscription policies — but the reality is that OpenClaw crossed into serious work mode, and once you can swap brains through a durable work layer, memory becomes the strategic layer that matters most. In this video, I share the inside scoop on what April's releases actually mean for builders: • Why OpenClaw is becoming a runtime abstraction for serious agentic work, not just a chatbot wrapper • How Anthropic's subscription changes and OpenAI's Codex access create opposite architecture assumptions • What makes a durable workflow survive model churn, pricing changes, and better local models • Where OpenBrain for OpenClaw fits and why memory can't live inside any one brain Leaders treating model choice as a permanent architectural decision are missing that the practical unlock is designing workflows that outlive a provider policy. Chapters 00:00 OpenClaw grew up in April 02:30 From viral demo to serious runtime 05:00 The boring stuff that makes work possible 07:30 Task flow, memory, and channel maturity 10:00 Anthropic's April move was deeply unpopular 12:30 OpenAI's opposite posture with Codex 15:00 Gemma 4 and the local model branch 17:30 Which model should handle this step 20:00 Durable workflows that survive the session 22:30 Memory can't live inside one brain 24:30 OpenBrain recipes for OpenClaw 25:30 Build the runtime so the model can change Subscribe for daily AI strategy and news. For deeper playbooks and analysis: https://natesnewsletter.substack.com/ Listen to this video as a podcast. - Spotify: https://open.spotify.com/show/0gkFdjd1wptEKJKLu9LbZ4 - Apple Podcasts: https://podcasts.apple.com/us/podcast/ai-news-strategy-daily-with-nate-b-jones/id1877109372
ADVERTISEMENT
I’m gonna have to start running your videos through an LLm to summarize them. I know that doesn’t help you monetize but it’s just so many words to get the content across my ADD can’t even
What about Hermes?
Great analysis. The thesis that memory becomes the strategic layer once models are swappable is exactly right, and I think most builders are underestimating how important that shift is. One thing worth adding: model swapping also breaks memory retrieval if your embeddings are tied to a specific model, so decoupling the embedding layer from the reasoning layer is a critical architectural step that doesn't get enough attention. The provenance point is also spot on — without clear labels on how memory was created, agent memory becomes liability rather than asset. I've been building along similar lines and this video validated a lot of those architectural choices. Looking forward to seeing the OpenBrain recipes evolve, especially the task flow worklog pattern.
Full Story w/ Open Brain Agent Memory: /p/openclaw-agent-runtime-model-swapping?r=1z4sm5&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true
The only thing I'd add to the model agnosticism, is that even with a robust harness/scaffolding in my experience there does seem to be a minimum viable reasoning threshold to make it really work well. I have a memory system I built that shares a number of similarities with your OpenBrain concept and have been running it for about 9 months now and I would say that the memory system is accessible by smaller MoE models with reasoning, it's not safe to assume that the smaller models will operationalize the memories the same way. I'd say DeepSeek v3.1 through v4 , GLM-5 and the latest "big" open source models behave comparably to GPT 5.5 and Opus 4.7 and Gemini 3.1 with their different flavors, but in my opinion once you drop below 30B parameters even with robust scaffolding I don't expect much, and it's exponential drops in performance on the way down to that floor. The other thing that was a pain point for me when allowing for model swaps is context window limits and compaction points; if you set a large global context window based on your biggest model you use, you're going to run into frustrating loops and freezes, and increased costs.
Bro like 5 years late with Openclaw 😂 it’s all about Hermes now
Prediction: every one of two videos he makes will be a hidden ad for his open brain thing.
The next big disrupter is going to be local semantic routing to various models. Several nascent projects are out there (vLLM, Sage, Manifest). That will be the final step in commoditizing inference. Improves responsiveness and cost. Self hosted cross system memory like OB1 is key too, but routing would dramatically cut costs and improve responsiveness. Kinda surprises me OpenAI doesn’t have an auto routing endpoint for a OpenClaw yet that’s probably coming too.
You can already do this with Pi out of the box…
I would love to see an up-to-date comparison with Hermes. Many of us switched from Open Claw to Hermes for its improved memory and reliability... Should I be switching back?
100% Agree. Big Like! This is turning the deal into you own the data and workflow. The AI model "Agent" is just a replaceable cog in your machine. The true Agency rests with you and your data and the various models are just cognitive workers. You keep the memory, the tasks are farmed out and soon can be done with local models easily and accurately. The cognitive utility vs the Agent as a service with your data is going to be a bitter pill for those who invested in these high value AI Clouds that will evaporate. We are at a time of rapid evolution and innovation on how the harness AI with competing approaches: - OpenClaw/Hermes/etc. - Karpathy Wiki - Your Open Brain - Jake Van Cleef Folders and Files
I love that you confirm my agent architecture is pretty spot on. I've been doing all this for the past month with my current project. I love the idea of Open Brain. I think the biggest issue is setup, then customization and prototyping harness structure. Not really accessible. Next is understanding what's going on under the hood. We may tell OpenClaw to set something up but it's not easy to see everything it did to get that part working. Also complications with potential conflicts between plugins/skills/tools etc. It's the difference between Lego and Meccano. Kids can set up Lego.
How does it survive daily updates?
That's of many reasons I use Pi over OC.
You talk too much. This should be a 5 min video
Your videos are always spot on and really well structured. Even though I’m familiar with these topics, they’re so information-dense that I still find myself rewinding to catch everything. Since your videos are already pretty long, I don't know if you could change much, but I’m seriously amazed you pull this off daily. I couldn't produce one video with this much info in a whole week. It’s wild how much knowledge you embed in every single one.
Your videos help me sort through the complex ever changing AI landscape. Thanks much!
and adult AI would be lovely as Im tired of changing the nappies on my Prime AIs
This is why I stuck with OpenClaw and have been putting in the work. I feel OpenClaw (or equivalent) is the right approach. By the way, my OpenClaw agent (the orchestrator) wants to use PostgreSQL + pgvector. It says this would best suit my business and support the goal of a local database with a local LLM installed to facilitate. I'll use Frontier models for the heavy lifting via API. All my data will be local and not in the cloud. Just need to pick the hardware.
What about HERMES! ? Can you make a video about that platform?